Add OpenCode plugin support alongside Claude Code

[kulterryan]

Summary

Adds first-class support for OpenCode alongside the existing Claude Code plugin. The skill content (protocols, personas, templates, Python validator) is shared between both hosts — OpenCode auto-discovers skills under .claude/skills/ as well as .opencode/skills/. Only the glue layer (commands + pre-declared agents + a host adapter) is OpenCode-specific.

Key properties

• Zero regressions to Claude Code. claude-plugin/, .claude/, .claude-plugin/, scripts/sync.sh, install.sh, and .github/ are bit-identical before/after.
• No duplicated skill content. One canonical copy of SKILL.md + references/.
• Thin, auditable glue. 22 new files, mostly short prose. No TypeScript plugin required — deferred to a follow-up.

What ships

New files

• .opencode/host-adapter.md — the single contract every OpenCode command reads alongside SKILL.md. Covers tool name map (Agent → task, WebSearch → websearch, etc.), pre-declared agent roster, parallel spawn pattern, Rule 13 runtime check, websearch caveat, context-file handling, publish chain overrides.
• .opencode/commands/*.md (10) — autodecision.md plus 9 hyphenated subcommands. OpenCode has no namespace/colon support, so commands are flat: /autodecision, /autodecision-quick, /autodecision-compare, /autodecision-revise, /autodecision-challenge, /autodecision-summarize, /autodecision-publish, /autodecision-plan, /autodecision-review, /autodecision-export.
• .opencode/agents/ad-*.md (9) — 5 persona agents (ad-optimist, ad-pessimist, ad-competitor, ad-regulator, ad-customer), plus ad-critique, ad-adversary, ad-sensitivity, ad-decide. All mode: subagent, hidden: true. OpenCode's Task tool requires pre-declared agent files (unlike Claude Code's inline-prompt Agent tool).
• scripts/sync-check-opencode.sh — structural validator. Confirms every command exists with valid frontmatter, every agent has mode: subagent + hidden: true, every command references the host adapter and the skill path, no orchestrator sets subtask: true.
• OPENCODE.md — host-specific reference doc (install, command map, how subagents work differently, known limitations, troubleshooting).

Modified files

• README.md — adds the OpenCode install section; dual command examples (colon vs hyphen) in Quick Start.
• CLAUDE.md — notes OpenCode install path under Installation; adds .opencode/ tree to Architecture diagram.

Key research findings that shaped this design

Capability	Claude Code	OpenCode	Port impact
Skills	SKILL.md + references/	Native skill tool; auto-discovers .claude/skills/ AND .opencode/skills/	Zero duplication needed
Parallel subagents	Task tool, inline prompts	task tool, but agents must be pre-declared as named files	9 pre-declared agents shipped
Slash commands	Colon namespace (/autodecision:quick)	No namespace / no subdirs — flat filenames only	Commands use hyphens
Plugin distribution	Marketplace + plugin.json	Local .opencode/ files OR npm JS-only plugin (npm cannot ship .md files)	Distribution: "copy .opencode/ into your project"
subtask: true	n/a	Forces command to run in child session	Not used on orchestrator — would block parallel spawning

Validator output

$ ./scripts/sync-check-opencode.sh
Checking .opencode/ tree...
  ok: all 10 commands
  ok: all 9 agents
  ok: host-adapter.md exists
  ok: skill discoverable at .claude/skills/autodecision/
  ok: validate-brief.py discoverable
sync-check-opencode: PASSED

Testing

• Validator passes
• Zero diffs to the Claude Code tree
• Live end-to-end test: run /autodecision "Should we cut pricing by 20%?" inside an OpenCode install and verify Phase 3 spawns all 5 persona subagents in parallel (deferred — requires OpenCode running on the machine)

Deferred (follow-up PR)

• Optional .opencode/plugin.ts with custom tools: autodecision_journal_append (schema-validated JSONL writes), autodecision_read_brief (slug resolver). Not required — the bash-based file I/O path works.
• Wiring sync-check-opencode.sh into .github/workflows/sync-check.yml.
• End-to-end smoke test automation.

Draft PR

This is a draft PR because the live end-to-end test against a real OpenCode install hasn't run yet. Once someone confirms /autodecision runs the full 10-phase loop inside OpenCode and spawns all 5 personas in parallel, this can be marked ready.

[harshilmathur]

Thanks for putting this together — the PR description is thorough, the "zero diffs to the Claude Code tree" discipline is exactly right, and the structural validator is a nice touch.

One bigger rework I'd want before this can merge, though. Right now .opencode/ is a hand-written parallel tree — 22 files maintained alongside claude-plugin/. Every future protocol change in claude-plugin/ would need matching edits here, and there's no content-level CI to catch drift (structural sync-check-opencode.sh only verifies files exist + have valid frontmatter).

What I'd want here:

1. scripts/sync-opencode.sh — treat claude-plugin/ as canonical and derive .opencode/ via sed rules: /autodecision:X → /autodecision-X, AskUserQuestion → question, Agent tool references → OpenCode task equivalent, plus a small Python patch for SKILL.md frontmatter if needed. Single source of truth, near-zero drift risk.
2. Wire it into CI — new job in .github/workflows/sync-check.yml that runs the script and fails on drift, mirroring the existing .claude/ invariant.
3. Keep hand-written only what genuinely has to be. The 9 pre-declared agents + host-adapter.md are new surface OpenCode requires (Claude Code embeds persona prompts inline in simulate.md, so those don't transform mechanically). Commands and SKILL.md should not be hand-written.
4. Guided installer — ./scripts/install.sh --opencode --global rather than "copy the directory manually."
5. Live end-to-end test before ready-for-review — non-negotiable. 2400 lines of unverified glue, even mostly prose, isn't something I can merge on trust. Phase 3 spawning 5 parallel persona subagents is the hard part — needs to be confirmed working in OpenCode.

Also: Codex support is out of scope here, but worth noting as a follow-up if OpenCode lands cleanly.

The overall direction (additive, non-invasive, skill content shared via autodiscovery) is sound. The rework is the execution model.